A real-time detection of battery pole before welding based on improved YOLOX

Rapid and accurate detection of the power battery pole area before welding is the prerequisite for accurately locating the welding starting point, and its performance determines the assembly efficiency and quality of the battery module. In view of the complex welding environment, low color contrast, and small area ratio, an improved model based on YOLOX is proposed. A multi-scale mixed attention module is introduced after the backbone feature to infer attention from different scales and enhance the key features of the target. A complex bidirectional fusion strategy that is conducive to small target detection is introduced at the neck of the model. The feature gradient flow is enriched through deep separable convolutions of multiple kernel sizes, and the original features are effectively utilized with the help of residual structures and multi-scale information is integrated. Experiments have shown that the optimized model is more stable to various environmental interferences while maintaining lightweight. The detection accuracy of the model is improved by 4.13% compared with the baseline model, the parameters are 6.27M, and the detection speed is 93 FPS. The overall performance is better than other models, providing an effective solution to the problem of welding initial point detection in power battery assembly.